15 years after Tom Herbert added skb->ooo_okay, only TCP transport
benefits from it.

We can support other transports directly from skb_set_owner_w().

If no other TX packet for this socket is in a host queue (qdisc, NIC queue)
there is no risk of self-inflicted reordering, we can set skb->ooo_okay.

This allows netdev_pick_tx() to choose a TX queue based on XPS settings,
instead of reusing the queue chosen at the time the first packet was sent
for connected sockets.

Tested:
  500 concurrent UDP_RR connected UDP flows, host with 32 TX queues,
  512 cpus, XPS setup.

  super_netperf 500 -t UDP_RR -H <host> -l 1000 -- -r 100,100 -Nn &

This patch saves between 10% and 20% of cycles, depending on how
process scheduler migrates threads among cpus.

Using following bpftrace script, we can see the effect on Qdisc/NIC tx queues
being better used (less cache line misses).

bpftrace -e '
k:__dev_queue_xmit { @start[cpu] = nsecs; }
kr:__dev_queue_xmit {
 if (@start[cpu]) {
    $delay = nsecs - @start[cpu];
    delete(@start[cpu]);
    @__dev_queue_xmit_ns = hist($delay);
 }
}
END { clear(@start); }'

Before:
@__dev_queue_xmit_ns:
[128, 256)             6 |                                                    |
[256, 512)        116283 |                                                    |
[512, 1K)        1888205 |@@@@@@@@@@@                                         |
[1K, 2K)         8106167 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@    |
[2K, 4K)         8699293 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[4K, 8K)         2600676 |@@@@@@@@@@@@@@@                                     |
[8K, 16K)         721688 |@@@@                                                |
[16K, 32K)        122995 |                                                    |
[32K, 64K)         10639 |                                                    |
[64K, 128K)          119 |                                                    |
[128K, 256K)           1 |                                                    |

After:
@__dev_queue_xmit_ns:
[128, 256)             3 |                                                    |
[256, 512)        651112 |@@                                                  |
[512, 1K)        8109938 |@@@@@@@@@@@@@@@@@@@@@@@@@@                          |
[1K, 2K)        16081031 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[2K, 4K)         2411692 |@@@@@@@                                             |
[4K, 8K)           98994 |                                                    |
[8K, 16K)           1536 |                                                    |
[16K, 32K)           587 |                                                    |
[32K, 64K)             2 |                                                    |

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
---
 net/core/sock.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index 542cfa16ee125f6c8487237c9040695d42794087..08ae20069b6d287745800710192396f76c8781b4 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -2694,6 +2694,8 @@ void __sock_wfree(struct sk_buff *skb)
 
 void skb_set_owner_w(struct sk_buff *skb, struct sock *sk)
 {
+	int old_wmem;
+
 	skb_orphan(skb);
 #ifdef CONFIG_INET
 	if (unlikely(!sk_fullsock(sk)))
@@ -2707,7 +2709,15 @@ void skb_set_owner_w(struct sk_buff *skb, struct sock *sk)
 	 * is enough to guarantee sk_free() won't free this sock until
 	 * all in-flight packets are completed
 	 */
-	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
+	__refcount_add(skb->truesize, &sk->sk_wmem_alloc, &old_wmem);
+
+	/* (old_wmem == SK_WMEM_ALLOC_BIAS) if no other TX packet for this socket
+	 * is in a host queue (qdisc, NIC queue).
+	 * Set skb->ooo_okay so that netdev_pick_tx() can choose a TX queue
+	 * based on XPS for better performance.
+	 * Otherwise clear ooo_okay to not risk Out Of Order delivery.
+	 */
+	skb->ooo_okay = (old_wmem == SK_WMEM_ALLOC_BIAS);
 }
 EXPORT_SYMBOL(skb_set_owner_w);
 
-- 
2.51.0.740.g6adb054d12-goog